Startup pace,full control.
Part 01 · Why a factory
The problem
Business needs predictability.AI has been giving it science projects.
Until now, business software has been built by hand, and AI projects most of all: every project a one-off, started from scratch. Many AI projects are experiments too, testing new designs and untried technology as they go. So the cost, the timeline and the result have been hard to predict.
Lessons from Toyota
Toyota didn't build better cars by hand.It built a better way to build them.
Car-making had the same problem until Toyota stopped treating every car as a one-off. We've built AI systems dozens of times, so we no longer start from scratch, and now we've built a software factory around that experience.
It runs on three of Toyota's ideas:
自働化
Automation with a human touch
Agents triage, plan, build, test and security-scan every change. Our engineers steer and approve.
See it in practice ↓部品共通化
Parts commonality
We bring components we've already built and battle-tested, so no project starts from zero.
See it in practice ↓改善
Continuous improvement
Every project starts from a library of proven blueprints, and every build improves them.
See it in practice ↓
Part 02 · How the factory works
How the factory works.
Each of Toyota's three ideas, in practice, then the security that runs through all of them.
2.1 · 自働化 in practice · The production line
Agents automate.Humans decide.
Automation with a human touch, in practice: every change runs the same seven-step line. Agents do the work, and at two of the steps an engineer decides.
- 01LinearIssue raisedEvery piece of work starts as an issue in Linear, our project tracker, so it is tracked from the first minute to the moment it ships.
- 02AgentTriageAn agent sizes the request, sets its priority and routes it to the right part of the line.
- 03AgentPlanAn agent drafts the build using our blueprints: a proven reference architecture, a standard tech stack and reusable components. No project starts from a blank page.
- 04EngineerHuman gateAn engineer reviews the plan and approves it before a line of code is written. Nothing gets built without a human decision.
- 05AgentBuildAgents write the code to the approved plan, using the standard components the blueprint calls for.
- 06AgentTest and scanAgents spin up isolated sandboxes (safe, throwaway copies of the system) to test every change and run the security checks (see below). Any failed check stops the line and sends the work back.
- 07EngineerPull requestOnly tested, security-scanned work reaches a pull request: the proposed change, ready for review. An independent AI reviewer checks it first; then an engineer reviews it and decides what ships.
Any failed check stops the line and sends the work back.
The method
The line runs on our method: how we build AI systems when an AI agent writes much of the code. It's published in full.
2.2 · 部品共通化 in practice · Standard components
We solve the hard problemsonce.
Parts commonality, in practice. Toyota doesn't reinvent the engine for every model, and neither do we. Most serious AI comes down to two hard problems: running a business process that spans agents, people and systems, and giving agents exactly the right context. We've invested in solving both once, so every project on the line starts with them already built.
Pharos · Process execution
Workflow for agents and people
Agent, human and system steps in one long-running flow, with state kept across days or weeks.
More on Pharos →
Alexandria · Context graph
Gives agents the right context
Knowledge as a graph, so agents get exactly the context they need.
More on Alexandria →
2.3 · 改善 in practice · The blueprint library
Built on blueprints,not experiments.
Continuous improvement, in practice. Every project starts from our library of blueprints. Toyota's rule is to do the work by hand first, understand its problems, then automate, so our blueprints come from systems we've built and run for clients: standard work, proven in the field.
改善 · Kaizen. Every project improves the library, with no radical, unproven overhauls.
What's in a blueprint
- Reference architectureA proven design for the problem at hand.
- Standard tech stackThe same well-understood foundations, every time.
- Reusable componentsPharos, Alexandria and more, ready to drop in.
- Test and security patternsThe checks the factory runs on every change.
- Model strategySmall models where they work; frontier models only where needed.
2.4 · Security on the line
Built to passyour security review.
Security isn't a review at the end. The factory automates a battery of security scans that run on every change, in order, and none can be skipped to hit a deadline. Each catches things the others miss, and an engineer approves the result.
- 01Secrets scanNo password, token or key gets into the code. Secrets live in a vault and are supplied at run time.
- 02Dependency and configuration scanEvery open-source component is checked for known vulnerabilities, and every cloud and infrastructure setting for risky defaults. Versions are pinned, brand-new releases are held back for a few days until they've proved safe, and each release has a bill of materials.
- 03Static analysisThe code is checked for known dangerous patterns, such as injection and broken authentication.
- 04AI security reviewAn AI reviewer reasons about how the parts of the code fit together, and finds the logic flaws pattern-matching misses.
- 05Human gateAn engineer approves the plan before any code is written, and reviews the result before it ships.
Agents with no keys to steal
In 2026 researchers tricked the coding agents of Anthropic, Google, GitHub and OpenAI into leaking secrets. Ours have none to leak.
- A sandbox for every agentAgents work inside isolated sandboxes: throwaway copies of the system, separate from anything that matters.
- Agents never hold a secretPasswords and keys stay outside the sandbox. When an agent calls a service, a proxy outside adds the credential on the way out, so even a hijacked agent has nothing to give away.
- Only approved destinationsA sandbox can reach only the services on its approved list. If something malicious ran inside, it couldn't send data anywhere unexpected.
For your security team
Because every check runs on every change, the evidence builds up as we work, and we hand it over.
- Scan reports for every release
- A software bill of materials
- The approval trail: who approved what, and when
- Architecture and data-flow notes
Model agnostic
We engineer every system to work with any AI model, so you're never tied to one provider. When your data or your regulator demands it, the same system can run on sovereign options: open models, in your own cloud and jurisdiction.
Part 03 · Working with the factory
You're inside the line,not waiting for updates.
When AI compresses build time, the bottleneck moves to coordination. Build fast with slow feedback and you just build the wrong thing faster. So your team works on the line with ours.
You're in the room from day one
A Slack Connect channel between your team and ours (or a free workspace we set up for you). You talk to our engineers and our agents in the same place, not through weekly status emails.
A shared, ring-fenced project space
A permissioned area of our Notion just for your project, which your team edits alongside ours. Our agents read it too, so they work to the same spec you can open and change.
Raise work from Slack, watch it in your portal
Message the triage agent in your channel and it files an issue on the line, marked for triage, so nothing is built until our team has scoped it. You follow progress in a secure client portal.
See every change land
Pull requests and merges post to the same Slack thread. Push back in the thread and the next iteration starts. The loop is hours, not weeks.
What comes off the line
What you getwith every release.
A working system
Running software you can test and use, deployed in your own cloud.
The code
Every line of source, in your own repositories, with its tests and documentation. Yours to keep, run and change, with no licence back to us.
Security scan results
The results of the full battery of checks every change passes: secrets, dependencies and configuration, static analysis and AI security review, each approved by an engineer.
What we check, and why →A software bill of materials
A complete list of every component in the system, its exact version and where it came from, so your security team can check it against new vulnerabilities as they emerge.
The AI Breakthrough Sprint
Find your AI breakthrough.
Not sure where to start? Start with a sprint, from £5,000 + VAT. If an AI idea won't pay off, you'll know by day 15.
Step 1 · One day
Discovery workshop
We uncover the hidden AI opportunities in your business and work out, with you, where to start: the one that's worth building first.
Step 2 · Two weeks
A working prototype in two weeks
Your idea goes onto the factory line for two weeks and comes off it as a working proof of concept your team can test, with a clear view of what production would take.
Now booking client slots on the factory line from January 2027.